Skip to content →

Progress of YamaX – DeepL2 –

Hi! It’s @coord_e.

We’ve been developed YamaX, our humanoid robot, for so long time.

After the discussion between @coord_e and @Nyanyan_Cube, We came to an agreement to continue the development of YamaX as a research.

The new project is called DeepL2.


We aim to develop automated learning that a humanoid walks.

At last, we would like to apply a result of learning in simulation to real YamaX.

Current status

We have implemented learning software in simulation environment.

However our simulation is not good at all.

It even cannot stand! Its legs is under the ground.

At first, we tried to make it learn without seeing what is happening in the simulation.

But it can’t be success, you know.

We were to use DQN(Deep Q Network), but memory was not enough with eight GPUs, so we use DDPG(Deep Deterministic Policy Gradient). We’re going to try A3C after we studied about parallel programming.

We use Gazebo as a simulator, ROS lunar as a framework, and Keras + keras-rl as a library to implement reinforcement learning in Python. To connect Python to gazebo, we use gym-gazebo. keras-rl seems to have stopped its development, we’re planning to transfer to use tensorforce.

Learning will be done with GPU in a docker container in EC2 GPU Instance, using nvidia-docker.


We haven’t implemented software for real YamaX.

Beause we don’t have any linux board to work on YamaX 4.0.

The white board was full of mistakes. I made a lot of modification but it doesn’t came to work, so I designed new board and ordered.

We’ll assembly it and use in YamaX 4.0.

I’ll write a post about this board later.






At last…

Anyway, we’ll keep up-to-date this blog.

I’m so sorry for stopping update of this blog.

Published in DeepL2 Y-modify YamaX Software Hardware Announcement Machine Learning apologize Management